Leveraging LLMs to Improve Neutrality and Reduce Fear in Spine MRI/CT Reports

Introduction: In the last decade, health care systems have been required to grant patients immediate access to electronic medical records upon request. While this improves transparency of care, viewing complex information such as radiology reports before appropriate medical follow-up may lead to unnecessary confusion and fear amongst patients. Towards this end, large language models (LLMs) provide a unique path forward in translating highly technical reports.

Methods: N=124 spine radiology reports, consisting of spine CT and MRI (including cervical, thoracic, lumbar, and total spine), were programmatically pooled from electronic health records for simplification using a GPT-4o model. Imaging indications spanned a broad range, from monitoring chronic disease (e.g., spinal metastasis) to evaluating acute surgical emergencies (e.g., cauda equina syndrome). The GPT-4o model was instructed to simplify radiology reports to a middle school reading level while preserving important clinical details. Model text simplification was conducted automatically using the AzureOpenAI toolkit. Natural language processing tools were used to assess readability and emotional tone. Sentiment analysis was conducted starting at the impression/conclusion sections of the original and simplified radiology reports.

Results: Text simplification was completed for all reports with an approximate run-time of 6 seconds per report. Post simplification, the reports significantly decreased from 12th to 6th grade reading level (Flesch-Kincaid; t(123)=50.2, p< 0.001, Cohen’s d=4.5). In addition, the texts were classified as more neutral (t(123)=15.6, p< 0.001, Cohen’s d=1.4) and less fearful (t(123)=13.7, p< 0.001, Cohen’s d=1.2). Post simplification, text neutrality increased by ~2x (0.31 to 0.57) and text fear decreased by ~4x (0.30 to 0.08).

Conclusion : This study demonstrates the feasibility and utility of LLMs in automatically simplifying spine radiology reports. Model simplification increased readability of medical information, and sentiment analysis revealed increased neutrality and decreased fear in the simplified reports. Our results provide a useful framework to enhance healthcare literacy in a patient-centered manner.