Generating colloquial radiology reports with large language models.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Radiology

Background:

• Patient access to medical records is increasing.
• Radiology reports contain complex medical jargon, hindering patient comprehension.
• Manual creation of patient-friendly reports is time-consuming for radiologists.

Purpose of the Study:

• To develop an automated method for translating radiology reports into accessible, patient-facing language.
• To leverage large language models (LLMs) with specialized prompts for colloquial report generation.
• To assess the accuracy and readability of LLM-generated patient-friendly reports.

Main Methods:

• A novel method using LLMs with specialized prompts was developed.
• Medical terms were automatically extracted and defined within the LLM prompt.
• 100 de-identified neuroradiology reports were translated into four reading levels.
• Translations were evaluated by radiologists for accuracy, likability, harm potential, and readability.

Main Results:

• The LLM approach achieved 88% accuracy for Findings and 93% for Impression sections at an 8th-grade reading level.
• Translations were 20% more accurate than a naive baseline method across all grade levels.
• Standard readability indices indicated improved readability compared to original reports.

Conclusions:

• LLM-generated translations at the 8th-grade level offer an optimal balance of accuracy and readability.
• This approach aligns with national health communication recommendations for patient-facing materials.
• Automated translation can enhance patient understanding of radiology reports without burdening radiologists.