Large language model reads radiologists' notes to flag patients for follow-up imaging
Researchers have successfully developed an artificial intelligence-powered tool capable of significantly improving the process of ensuring patients adhere to follow-up imaging recommendations.
In fact, use of the model dramatically strengthened one health system's internal flagging processes, according to a new study published in NEJM Catalyst. Experts involved in the analysis believe adopting such systems in clinical care could improve diagnostic accuracy and patient outcomes.
“Missed opportunities for diagnosis are a critical subset of diagnostic errors that can lead to adverse patient outcomes," George Oliver, MD, PhD, VP of informatics for the Parkland Center for Clinical Innovation, Dallas, and colleagues noted. "These errors frequently arise from failures in the diagnostic process, particularly in ensuring that recommended follow-ups are scheduled and completed."
Parkland Health conducts over 500,000 radiologist studies annually, but has faced challenges relying on structured notes templates, or macros, in EHRs to pinpoint patients requiring follow-ups. The team sought to develop a tool that could better flag these instances using only rads' notes. They built a custom artificial intelligence agent that uses a pre-trained large language model (LLM) capable of reviewing clinical impressions and extracting and standardizing key details from notes to determine the need for follow-up. The LLM was integrated into the health system’s EHR so it could transmit information pertaining to the need for additional imaging once recommendations are identified in radiology reports.
Researchers initially tested the model on a random sample of 10,000 radiologist notes before expanding their analysis by another three months and 120,000 imaging studies. The team observed significant improvements in follow-up adherence, accurately identifying 97% of rad recommendations and correctly flagging 6.18 times more cases than the organization's previous macro-based system (83 vs. 513). AI also yielded 94% accuracy in characterizing the timing of when follow-up exams needed to be completed, the type of exam and the diagnosis that prompted the reading radiologist to suggest additional imaging.
The LLM’s solid performance enabled staff to more easily identify patients in need of follow-ups, resulting in more scheduled scans.
“By enhancing the reliability of follow-up identification and standardizing key details, this approach increases the likelihood that patients receive appropriate care with the intention of optimizing healthcare outcomes in high-volume clinical settings,” the authors concluded.
Read more here.
In addition to her background in journalism, Hannah also has patient-facing experience in clinical settings, having spent more than 12 years working as a registered rad tech. She began covering the medical imaging industry for Innovate Healthcare in 2021.