The growing volume of imaging in the Yale New Haven healthcare system has placed increasing demands on radiologists and staff. Many institutions are facing a similar challenge.
THE PROBLEM
The healthcare system began looking at artificial intelligence technology in the context of emergency department workflows. Part of this need was figuring out the best way to prioritize patients coming in with acute findings discovered during head CT scans.
“We participated largely to start operational and research interests,” said Dr. Melissa A. Davis, associate professor of radiology and biomedical imaging and vice chair for imaging informatics, radiology and biomedical imaging at Yale New Haven. “These technologies are still largely new, so their impact is not yet fully known. We wanted to be at the forefront of that conversation.”
PROPOSAL
The original proposal was that the AI would evaluate non-contrast head CTs at the main hospital and the health care system would validate the findings with dedicated neuroradiology-trained radiologists.
“We would also evaluate whether there were any gains in turnaround time for these images and report the discrepancy rates for a retrospective cohort,” Davis noted. “The plan was to deploy solutions that would identify and prioritize intracranial hemorrhages in non-contrast CT examinations.
“The biggest expectation was around the user experience,” she added. “It couldn’t be a new step in the workflow. The expectation was that everything would be seamless and integrated into the radiologist’s current workflow. Without that approach, technology adoption would be a challenge and a burden.”
MEETING THE CHALLENGE
The AI program is led by the radiology and biomedical imaging department and is aligned with IT to ensure smooth integration and transition, Davis explains.
“In the beginning, the AI was deployed for the emergency radiologists and neuroradiologists,” she said. “It was used to evaluate CT scans of the head for potential. When a CT scan of the head was performed, the AI tool analyzed the images and flagged any cases with findings suspicious of blood.
“Integration was a key aspect of this implementation,” she continued. “The solution needed to integrate seamlessly with existing radiology workflow systems. Embedding an icon in the worklist to indicate when AI had flagged a case enabled easy adoption and use of the technology.”
“The adoption of AI technology in healthcare can lead to significant improvements in efficiency, patient care and outcomes.”
Dr. Melissa A. Davis, Yale New Haven
Additionally, there was a notification that would appear when an acute finding was detected, allowing radiologists and other physicians to be immediately notified.
“There was a significant reduction in turnaround time for our level 1 trauma center ED, but not at other locations,” Davis reported. “It did identify an outpatient with a bleed in the head who would otherwise have been sent home. This was the biggest initial success. We also noticed that radiologists began to see some comfort in having a second set of ‘eyes’ on these cases.
“As implementation evolved, the use of AI expanded from head CT scans to other radiology applications, including the detection of pulmonary embolism and coronary artery calcification,” she continued. “This also expanded the conversation about AI’s place in radiology from one focused on adoption to one focused on expanding our current workflows.”
While the staff often talks about buying time, Davis said the AI sometimes even slows her down a bit, forcing her to look much more closely at an image or area — especially if she disagrees with the AI.
“You always have to think critically,” she said.
RESULTS
“For Yale, there are three potential areas of downstream value that we looked at: increased sensitivity and specificity that leads to improved accuracy; discovery of incidental findings that lead to more clinically appropriate interventions; and improved efficiency that helps shorten the length of stay of the patient.” Davis said. “Studies have highlighted these consequences.”
In one study of improved accuracy pointed out by Davis, an AI algorithm was applied to a retrospective cohort of 1,387 consecutive CT pulmonary angiograms. The prevalence of pulmonary embolism (PE) was 13.6% (189 cases). The algorithm was 93% sensitive and 96% specific in detecting PE. The positive predictive value was 77% and the negative predictive value was 99%.
Davis said the conclusion to be drawn from the study is that the high negative predictive value demonstrates success in screening.
ADVICE FOR OTHERS
For healthcare organizations considering the adoption of similar AI technology in their clinical workflows, there are several crucial pieces of advice to consider, Davis said.
“First, it is essential to thoroughly assess your organization’s specific clinical needs and challenges,” she said. “AI technology can be a powerful tool, but its effectiveness depends on how well it fits your goals.
“Identify areas where AI could potentially have a meaningful impact, such as improving workflow efficiency, improving diagnostic accuracy, or accelerating communication of critical results,” she continued. “Prioritize AI adoption in these areas as it can be challenging to implement across the board at the same time.”
Second, involve key stakeholders, including healthcare professionals, IT teams and administrators, in the decision-making process, she advised.
“Make sure there is buy-in and support from all relevant parties,” she emphasizes. “Healthcare providers who will be using the technology directly should be involved in the selection process as their feedback and insights will be invaluable to successful implementation.
“Third, carefully evaluate AI technology vendors,” she continued. “Consider factors such as the accuracy of the AI algorithms, the ease of integration with existing systems, the reputation of suppliers and the level of ongoing support and updates. Look for references and case studies from organizations that have successfully implemented similar AI solutions to gain insight into their experiences.”
Further invest in robust IT infrastructure and ensure the organization’s systems can seamlessly support AI integration, she said.
“Collaboration between your IT team and the AI vendor is crucial to address any technical challenges that may arise during implementation,” she added.
Finally, focus on education and training, she said.
“Ensure that healthcare professionals, especially radiologists and physicians, are adequately trained to use AI technology effectively,” she advised. “Emphasize the importance of maintaining a critical mindset when working with AI and encourage continuous learning to stay abreast of the latest developments.
“The adoption of AI technology in healthcare can lead to significant improvements in efficiency, patient care and outcomes,” she concluded. “However, successful implementation requires careful planning, stakeholder engagement, vendor evaluation, IT infrastructure preparedness and ongoing education. By taking these steps, healthcare organizations can leverage the potential of AI to improve their clinical workflows and ultimately provide better care to their patients. “
Follow Bill’s HIT coverage on LinkedIn: Bill Siwicki
Email him: bsiwicki@himss.org
Healthcare IT News is a HIMSS Media publication.