'RadGPT' system improves radiology report readability and is ready for 'immediate' use
A specialized large language model-based system developed specifically for radiology applications could be key to helping patients better understand their imaging reports.
The 21st Century Cures Act requires medical practices to give patients immediate access to their radiology reports. Though this is logistically convenient for patients, radiology reports are full of complex medical jargon that the average individual might not understand. This can cause problems for patients, especially when providers are not immediately available to explain the results.
“On one hand, immediate access to health information promotes greater transparency and allows patients to be more involved in and informed about their medical care. Studies suggest that most people prefer instant access to their own health records, including radiology reports,” Sanna E. Herwald, MD, PhD, with the department of radiology at Stanford Medicine, and colleagues noted. “However, access to such information prior to explanation by a healthcare professional can lead to misinterpretation, anxiety and possible harm.”
Large language models have shown great potential for translating complex information into lay language that is more easily understood. Experts in the medical field are especially interested in the utility of LLMs to do this with medical information, but many have cautioned that it must be done carefully, as the performance of LLMs in healthcare largely depends on the data used to train it.
With this in mind, researchers designed an LLM-based system with patients’ radiology reports in mind. Enter “RadGPT.”
The model was trained on radiology report impressions generated between 2012 and 2020. The data from the reports were used to create concept-based patient explanations and concept-based question-and-answer pairs. Researchers also generated report-based question-and-answer pairs directly from the report impressions using an LLM without concept extraction. One radiologist and four residents rated the results using a standardized rubric.
Nearly all the RadGPT materials were scored highly by the readers, with 50% earning the highest possibly ranking. None of the answers contained information that could be misconstrued by patients in a way that would affect their safety or care—an important finding, as this is the primary concern among experts about deploying LLMs in patient settings.
RadGPT performed best with concept-based, LLM-generated questions. However, its report-level, LLM-based questions also impressed raters, too, with 92% earning the highest possible ranking.
The system performed so well, that the group suggested it is ready for immediate use in electronic health records.
“RadGPT offers tailored, patient-centered educational material, both about the overall radiology report as well as specific concepts within the report, that poses a low risk to patient safety and in the vast majority of cases is ready for immediate use without further human curation,” the authors wrote. “RadGPT content appears poised to fulfill the critical need for real-time, accurate, personalized and patient-centered educational materials, which would ultimately allow patients to access explanations through hyperlinks to any or all of the concepts mentioned in their reports.”
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.