Comprehensive Summary
This pilot study evaluates the ability of GPT-4, a Large Language Model (LLM), to translate lumbar spine imaging reports into patient-friendly language. 93 imaging reports from spinal surgery patients were individually pasted into GPT-4 with a simple prompt for plain-language translation. The simplified outputs were evaluated by an orthopedic surgeon for accuracy and completeness, rated based on these parameters using a 1-5 Likert scale, and analyzed for readability using the Flesch-Kincaid scale. Translations were similar in word-count to the original reports, while reading level was simplified from grade 11-12 to grade 8-9. However, the initial outputs had frequent omissions or inaccuracies: 35% had major omissions, 20% minor omissions, 6% major inaccuracies, and 3% minor inaccuracies. 36 errorless reports were used to create an enhanced prompt for reevaluation of the remaining reports. While this improved omission rates (major 7%, minor 4%) and Likert scores (3.48 to 4.39), inaccuracy rates remained consistent. These findings suggest LLMs can simplify language in lumbar-spine reports, but errors in accuracy and completeness remain significant.
Outcomes and Implications
The use of LLMs for translation of spinal imaging reports into plain language has potential to make medical information more accessible to patients. This allows patients to have a deeper understanding of their spinal pathology and greater involvement in treatment decisions. Additionally, LLMs can reduce the burden of translating complex imaging reports, allowing physicians to spend more time on providing care and clinical decision-making. Despite the potential shown by GPT-4 in this setting, persistent inaccuracies and omissions highlight the need for physician oversight to ensure reliability in early-stage clinical use.